Just-in-Time Learning

In the early 1990s, "just in time learning" was a trendy topic among cognitive scientists, educators, and corporate trainers. The idea was that you can't teach people everything they need to know in the classroom, and that some knowledge would need to be elaborated, adapted, or even corrected in practical situations.

Among applied researchers like myself at the Institute for Research on Learning (Menlo Park, CA), just-in-time learning was a philosophically welcome bandwagon, as the idea provided a nice lead- in to our theories of knowledge and learning. These theories were variously referred to as "constructivist" or "constructionism" -- or our preferred term, "situated learning" -- and dated to early 20th century work by John Dewey, Vygotsky (a Soviet psychologist), Piaget (a Swiss educational psychologist) and many others.

"Situated" theories of learning have been supported and elaborated along a wide range of analysis, from neuropsychology (conceptualization accomplishes higher-order perceptual-motor coordination), to cognitive psychology (human memory is process- oriented, facts and procedures are not literally stored in the brain as in a digital computer), to social science (our understanding of how to behave is coupled to our conception of roles in activities, i.e., identities).

According to these theories, just-in-time learning, although suggesting a kind of fortunate catch or recovery, might be interpreted not as making up for poor training, but inherent in all human action. At root, it means that everything you is on many levels of analysis improvised, adapted, and activated on the spot. Even when you recite a poem, you cannot exactly replicate a past performance. Intonation and thus meaning will change, new implications might occur to you. After many repetitions, some words might even begin to appear strange.

In work practice, of course, adaptations are often required -- procedures don't fit a situation, parts may require realignment, subproblems arise, workarounds are invented and themselves adapted.

Now here lies the nexus of a dilemma for trainers, managers, and computer scientists: Most automation is based on a non-improvisatory theory of performance. The very essence of the advanced software we develop for automated life support, navigation, and spacecraft control is based on a simple notion that before any action occurs, there is a program to dictate what to do. In our technical designs, this appears as an "architecture" involving two steps: PLAN and EXECUTE. For example, programmers working on robotic systems might specialize in one of these steps.

A busy area of research for 20 years in artificial intelligence has concerned how to deal with the reality that the future cannot be strictly predicted and so plans go awry. Thus methods have been developed for adapting plans during execution, as well as creating plans that have many complex alternative paths. Such advanced software should be contrasted to a remark made at the Mars Exploration Rover Preliminary Design Review in early 2001, that there would be no conditionals in MER plans. This meant that the rovers' operation would be strictly predictable; there would be no place for robot adaptation, let alone just-in-time learning (or replanning). People would decide and control all robotic actions. (This constraint was loosened somewhat as the engineers and managers gained confidence in the onboard systems, especially for navigation using obstacle avoidance.)

So where does this leave attempts to improve crew self-reliance on future space flights? Obviously, training, simulation, and reliable systems will go a long way -- knowledgeable people and robust automation will make complex operations more routine. The question is what training and tools can we provide to handle the unexpected situations and the complications where multiple things go wrong at once?

Although we can sit in our offices back at Ames and imagine problems and solutions, MDRS provides new insights and perspectives on what it's like to rely on a complex system that you do not fully understand and yet must repair. We experience stressful situations analogous to what astronauts might encounter.

Consider what happened yesterday with the power system. I counted seven interacting components -- a gas generator, a cable/switch, the inverter, the power load (AC in the hab, DC in the GreenHab), two banks of batteries, and an external charging system (the vehicle we connected to the hab batteries). To make this system work with a new generator, we had to change the setting of every component! This included physical changes (e.g., resplicing the cable) and programmatic changes (e.g., inverter-generator thresholds).

Along the way, we had to understand the interactions and how systems operated. For example, why was the upper bank of batteries always more depleted than the lower bank? Was this related to the GreenHab's DC pumps? What happened to the power provided by the car when the car got hot and it automatically engaged the fan? Why did throttling the generator cause the inverter's reading of generator voltage to start stabilizing? What would happen to hab appliances if we settled on 55 Hz 130 volt operation? We considered dozens of interacting variables before we found a configuration that worked. And we inspected or tested as many physical places visually and with voltmeters to discover and identify these variables.

The three-hour troubleshooting process was documented with the crew tracking system (video, audio, & location), by video recording our work at the inverter and generator, and by the agent system's log of OneMeter data about volts and amps throughout the hab's power system. Analyzing this data will take several days and might be a good topic for a research paper.

One of the topics of the paper might be just-in-time learning -- what and how we learned about the power system -- particularly the operation and interaction of the seven components -- during those three hours. And regarding our agent design, we will then inquire, what assistance by agents would have facilitated and accelerated this learning process?

Bill Clancey
Commander, MDRS Crew 49